But "search" and "getting an answer to a question" are two different things, aren't they? I realize that the trend has been going this way for a long time - probably since Ask Jeeves started blurring the line - and this is indeed how a lot of people try / want to use search engines, but still... I wish that Google (and competitors) would have separate pages for something like "Ask Google" vs. traditional search (where I want to find a particular document or quality content on a certain topic instead of just getting a specific answer).
May I ask how old you are? I'm 38 and I've been trying hard to break my 10 year-old of the habit of just typing questions into search engines (or telling me to "Ask Google" whenever she asks me a question and I say, "Oh, I don't know").
They're perfectly capable of implementing all the same search operators as 1990s Yahoo and 2000s Google. It's a solved problem.
The issue is that they don't want to. They'd rather be a middleman offering you "useful recommendations" (that they or may not sell to the highest bidder) instead of offering you value.
Agreed. So many times I have to put Wikipedia or Reddit behind my search to get anything useful out of google. So it can work. Google is clearly prioritizing junk over value
I've been using it for a while now. It is marginally better, but not exactly night and day. It seems to struggle with intent at times, and others I just get the same bland results as the free engines. The privacy is a big plus however.
Probably because it costs money and it also likely also will quickly succumb to sloppification by experimenting with their own ai and having an unstable founder…
I'm using it for about two years and i haven't seen any sloppification. I see it as a feature that it is a paid service because i hope it will be a sustainable model for them to keep it as it is. I think it's a no brainer to pay for it instead of all the suffering people describe here. The founder remark i don't get
Right now Kagi is a better search engine than Google. Why should some eventual demise in the future discourage anyone? There is no cost of switching and you can start using it right away
Would recommend to just try their 100 free searches.
Their results are good, but it’s hard to have an objective measure. For me, it’s the little features that make it worth it (and that they have a forum for feature requests, and a proper changelog).
I've been disappointed with pretty much all recent SEs (DDG being among the very worst). Having been an early Scroogle user, ixquick (startpage) and a few other ones I dearly miss, I've been using https://freespoke.com/ lately and find it tolerable.
I was using searx instances with reasonable results but many of them started failing recently.
Anyway, I hope everyone finds a good one. I fear things will only get worse though.
The biggest reason SEO is profitable is because low quality sites run display ads. That is the lifeblood, and intrinsic motivation, for these sites to even exist.
Google operates the largest display ads network. They literally *pay* websites for SEO spam, and take a very healthy cut off the top.
I wish people would stop acting like Google has been in a noble battle against the spam sites, when those sites generate Google billions of dollars a year in revenue.
The obvious question is, why would they ruin search for display? The answer is greed combined with hubris. They were able to double dip for years, but they killed the golden goose.
Everybody with a brain knew this would happen when they bought Doubleclick, and it took longer than expected, but here we are.
I think the search and ad code base may not be explicitly co-mingled, but they are implicitly co-mingled. And return worse search results than the early 2000s code base.
Both are explicitly related to the web at large. Google sells ads on more than two million sites, and those mostly aim to be the kind of sites that feature in search results. I'd say that the two code bases are related, by virtue of operating on the same data structure.
You remember the pagerank paper? It described how Google classifies each page on a scale from "something that links to good pages" to "informative page". Google and other search engines produces links to the latter. And since then, web site operators have had strong incentives to be on the "informative page" end of the range. Today I don't think the 2000s code can find a lot of pages of the former kind (well, outside Facebook).
It produced very nice results back then. It was/is good code, but good results need more than just good code, it needs good input too.
> The LD50 of caffeine in humans is dependent on individual sensitivity, but is estimated to be 150–200 milligrams per kilogram of body mass (75–100 cups of coffee for a 70 kilogram adult).
I think even when the answer is "right" in some sense, it should probably come within the context of a bunch of caveats, explanations, etc.
But maybe I'm just weird. Oftentimes when my wife or kids ask me a question, I take a deep breath and start to say something like "I know what you're asking, but there's not a simple or straightforward answer; it's important to first understand ____ or define ____..." by which time they get frustrated with me.
> I think even when the answer is "right" in some sense, it should probably come within the context of a bunch of caveats, explanations, etc.
Funnily enough, this is exactly what the LLM does with these questions. So well that people usually try to tweak their prompts so they don't have to wade through additional info, context, hedging, and caveats.
If you’re aiming to lose weight safely the rule of thumb is 3 lbs a week. 3000kcal per pound works out to an average deficit of about 1280 calories per day. Max.
Obese people can lose a bit more under doctor supervision. My understanding is that it’s tied partially to % of body weight lost per week and partly to what your organs can process, which does not increase with body mass.
I don’t think absolute numbers are very useful here. You need around 5–10% reduction in calorie intake to get any weight-loss effect going, and I wouldn’t reduce by more than 20% (relative to weight-maintaining intake — it’s different if you’ve been seriously overeating) if you want it to be sustainable longer-term.
So for example if your weight is stable at 2500 kcal per day, I would start by reducing the intake by 250–500 kcal, but not more. If this works well for a month or two and then you want to lose weight faster, you can still reduce your intake further. You generally have to do that anyway even just to maintain the velocity, because weight loss also tends to reduce calorie expenditure.
First and foremost, you need to monitor your calorie intake against weight. Here is a useful text about that: https://www.fourmilab.ch/hackdiet/
Your body will get more efficient at whatever exercise you do to make the calories work out. So over time you’ll either have to increase your exercise or rein in the calories a bit more to achieve a sustained result.
That is assuming that you do any exercise. But yes, and the method I linked to explains how to handle any such variation in calorie expenditure regardless of its cause.
I don't see the problem with the answer, and the question is already garbage. Plus, the LLM hedges its advice with precautions.
I get a pretty good summary when I paste the question into Google. It comes up with a ballpark but also gives precautions and info on how to estimate what caloric restriction makes sense for you within the first 3 sentences.
And all in a format someone is likely to read instead of clicking on some verbose search result that only answers the question if they read a whole article which they aren't going to do.
This seems like really lame nit picking. And I don't think it passes the "compared to what?" test.
The basic problem is it says reduce "by 1,500 - 1,800" rather than "to 1,500 - 1,800" (not that that answer is much better). Yes, it's a garbage question, but the first answer is unsafe in all circumstances. The simplest solution here is to show nothing.
Yes. You can't just say that someone eating 1800 over the recommended 2000 will perpetually gain weight. Weight maintenance calories will depend on the weight of the person.
A 500lbs man will need to consume 4000kcals/day to not lose weight. Cutting 1800 of that is realistic and might be good advice on the LLM's part, so it really depends on how GP asked the question.
Funny thing is you can train a small BERT model to detect queries that are in categories that aren’t ready for AI “answers” with like .00000001% of the energy of an LLM.
That's (obviously) a bit of an exaggeration. BERT is just another transformer architecture. Cut down from ~100 layers to 1, ~1k dimensions to ~10, and ~10k tokens to 100, and you're only 1e6 faster / more efficient, still a factor of 10k greater than your estimate and also too small to handle the detection you're describing with any reasonable degree of accuracy.
I literally have DistilBERT models that can do this exact task in ~14ms on an NVIDIA A6000. I don’t know the precise performance per watt, but it’s really fucking low.
I use LLM to help with training data as they are great at zero shot, but after the training corpora is built a small, well trained, model will smoke an LLM in classification accuracy and are way faster - which means you can get scale and low carbon cost.
In my personal opinion there is a moral imperative to use the most efficient models possible at every step in a system design. LLM are one type of architecture and while they do a lot well, you can use a variety of energy efficient techniques to do discrete tasks much better.
Thanks for providing a concrete model to work with. Compared to GPT3.5, the number you're looking for is ~0.04%. I pointed out the napkin math because 0.00000001% was so obviously wrong even at a glance that it was hurting your claim.
And, yes, purpose-built models definitely have their place even with the advent of LLMs. I'm happy to see more people working on that sort of thing.
reduce their daily calorie intake to 1,500 - 1,800 calories
or
reduce their daily calorie intake by 1,500 - 1,800 calories
These are very different answers, unless you’re consuming ~3,300 calories per day. These kinds of ‘subtle’ phrasing issue often results in AI mistake as both words are commonly used in advice but the context is really important.
Oh yeah! No, reduce to not reduce by. Though at the time I was eating a few things that had high calories that I didn’t realize so it would have been the same.
Your trainer advises you to reduce your calorie intake to between 200 and 500 calories per day? [0] That sounds very, very hazardous for anything other than very short term use, and (given the body's inbuilt "starvation mode") probably counterproductive, even then.
[0] Note that the robot suggested to reduce calorie intake by 1,500->1,800 calories, and the recommended calorie intake is 2,000.
People losing weight are probably eating more than 2000 per day to begin with. But if you go from 2800 down to 1500 you’re already likely to exceed 3 lbs of weight loss per week that is recommended without doctor supervision. If you need to lose more than 150 lbs in a year because you’re well past morbid obesity then you need staff, not just a food plan.
For whatever it’s worth, in response to the same question posed by me (“what is the ld50 of caffeine”), Google’s AI properly reported it as 150-200 mg/kg.
I asked this about 1 minute after you posted your comment. Perhaps it learned of and corrected its mistake in that short span of time, perhaps it reports differently on every occasion, or perhaps it thought you were a rat :)
The median lethal dose (LD50) of caffeine in humans is estimated to be 150–200 milligrams per kilogram of body mass. However, the lethal dose can vary depending on a person's sensitivity to caffeine, and can be as low as 57 milligrams per kilogram.
Route of administration
Oral 367.7 mg/kg bw
Dermal >2000 mg/kg bw
Inhalation LC50 combined: ca. 4.94 mg/L
That’s half the people in a caffeine chugging contest falling over dead. The first 911 call would be much much earlier. I doubt you’d get to 57 mg before someone thought they were having a heart attack (angina).
I also got similar and just tried, we are posting within minutes.
--
The median lethal dose (LD50) of caffeine in humans is estimated to be 150–200 milligrams per kilogram of body mass. However, the lethal dose can vary depending on a person's sensitivity to caffeine, and can be as low as 57 milligrams per kilogram.
Route of administration
LD50
Oral
367.7 mg/kg bw
Dermal
2000 mg/kg bw
Inhalation
LC50 combined: ca. 4.94 mg/L
The FDA estimates that toxic effects, such as seizures, can occur after consuming around 1,200 milligrams of caffeine.
Is this really true? The linear algebra is deterministic, although maybe there is some chaotic behavior with floating point handling. The non deterministic part mostly comes from intentionally added randomness, which can be turned off right?
Maybe the argument is that if you turn off the randomness you don’t have an LLM like result any more?
Floats are deterministic too (this winds up being helpful if you want to do something like test an algorithm on every single float); you just might get different deterministic outcomes on different compilation targets or with threaded intermediate values.
The argument is, as you suggest, that without randomness you don't have an LLM-like result any more. You _can_ use the most likely token every time, or beam search, or any number of other strategies to try to tease out an answer. Doing so gives you a completely different result distribution, and it's not even guaranteed to give a "likely" output (imagine, e.g., a string of tokens that are all 10% likely for any greedy choice, vs a different string where the first is 9% and the remainder are 90% -- with a 10-token answer the second option is 387 million times more likely with random sampling but will never happen with a simple deterministic strategy, and you can tweak the example slightly to keep beam search and similar from finding good results).
That brings up an interesting UI/UX question.
Suppose (as a simplified example) that you have a simple yes/no question and only know the answer probabilistically, something like "will it rain tomorrow" with an appropriate answer being "yes" 60% of the time and "no" 40%. Do you try to lengthen the answer to include that uncertainty? Do you respond "yes" always? 60% of the time? To 60% of the users and then deterministically for a period of time for each user to prevent flip-flopping answers?
The LD50 question is just a more complicated version of that conundrum. The model isn't quite sure. The question forces its hand a bit in terms of the classes of answers. What should its result distribution be?
I get your point wasn't this specific example, it's perhaps not a very good example of being dangerous: Getting that much caffeine into your bloodstream takes quite a commitment, and someone who knows the term LD50 is perhaps not very likely to think it indicates what is safe to consume. It's also not something you're likely to do accidentally because you've looked it up online and decided to test it.
In the most concentrated form in typical commercial caffeine tablets, it's half to one fistful. In high-caffeine pre-workout supplements, it's still a quantity that you'd find almost impossible to get down and keep down... E.g. a large tumbler full of powder of mine with just enough water to make it a thick slurry you'd likely vomit up long before much would make it into your bloodstream...
I'm not saying it's impossible to overdose on caffeinated drinks, because some do, and you can run into health problems before that, but I don't think that error is likely to be very high on the list of dangerous advice.
Hmm my search returns “between 150 to 200 mg per kilogram”, which is maybe more correct?
Also, in what context is this dangerous? To reach dangerous levels one would have to drink well over 100 cups of coffee in a sitting, something remarkably hard to do.
as ever, machine learning is not really suitable for safety/security critical systems / use cases without additional non-ML measures. it hasn’t been in the past, and i’ve seen zero evidence recently to back up any claim that it is.
I don't doubt the news article on this, but even with caffeine pills/powder it's near half a fistful to get to LD50 judging by my caffeine tablets. It's not impossible to consume, but it'd be distinctly unpleasant long before you get even anywhere close to dangerous levels.
For my high-caffeine pre-workout powder, I suspect I'd vomit long before I'd get anywhere near. Pure caffeine is less unpleasant, but still pretty awful, which I guess is why we don't see more deaths from it despite the widespread use.
I agree with you that there really ought to be caution around giving advice on safety-critical things, but this one really is right up there in freak accident territory, in the intersection of somewhat dangerous substances sold in a poorly regulated form (e.g. there's little reason for these to be sold as bulk powders instead of pressed into pills other than making people feel more macho downing awful tasting drinks instead of taking pills).
I wonder if they’re thinking 200mg per kilo to trigger cell death. I have trouble believing a human heart surviving a dose of 50mg/kg. Half of them surviving four times that much? No. I don’t believe it.
Found an article about a teenager who died after three strong beverages. The coroner is careful to point out that this was likely an underlying medical condition not the caffeine. The health professional they interviewed claims 10g is lethal for “most” people, which would be 100-150mg/kg. That still seems like something an ER doctor would roll their eyes at.
Your example doesn't interact with the chicken littling in this thread.
> The hearing was told the scales Mr Mansfield had used to measure the powder had a weighing range from two to 5,000 grams, whereas he was attempting to weigh a recommended dose of 60-300mg.
Nothing to do with an LLM nor with someone not knowing the exact LD50 of caffeine. Just "this article contains someone dying of caffeine overdose, and we're talking about caffeine overdose here, therefore LLM is dangerous."
> some people use caffeine powder / pills for gym stuff apparently.
At 200mg per pill, which is the strongest I had, I'd still have to down some 70+ pills in one go. Not strictly impossible, but not something you could possibly do by accident, and even for the purpose of early check-out, it wouldn't be my first choice.
An accident with it in powdered form is possible - people who use them are often used to pre-workout supplements tasting awful, and so might be prepared to down it as fast as possible - but it's a big enough volume of powder that it really is a freak accident.
And if on purpose, using caffeine would just be staggeringly awful...
the problem isn’t someone’s intent (on purpose/by accident).
it’s intent (want to improve my gym performance so down a bunch of caffeine) combined with incorrect information gained from what is supposedly a trustworthy source (the limit presented is much higher than it actually is for humans).
If they're searching for LD50, they're already setting themselves up for errors, even with the right information. The LD50 isn't a safe dose, after all, but the mean lethal dose. While it's not great if its wrong, if people search for an LD50 thinking it indicates what they can safely take, it's already going to be hard to protect them against themselves.
Even that seems high. I don’t feel good with 200mg per human, not per kilo. I can’t imagine drinking ten times as much and not being in the ER. A hundred times that much? No fucking way.
Yes, Google's AI chatbot confidently claimed yesterday that US passports have a fingerprinting requirement, which is absolutely not true. These things can't be trusted to emit even basic facts without somehow screwing it up and it's frankly depressing how they are worming their way into almost everything. I hope this particular hype train is derailed as soon as possible.
> I've been trying hard to break my 10 year-old of the habit of just typing questions into search engines
Honest question: why?
I understand not wanting to use Google (the search engine) or not wanting to support Google (the company). But I don't see with the issue with just looking up questions.
I'm 10 years younger than you, and I've been reaching for search engines first since I was 7, I think. Basically since I learned how to turn the computer on and open a web browser.
Right, A lot of times I'm searching for a filing. Or a site link. I do not ask questions when I'm doing so, that's ridiculous. I don't ask questions if I'm searching for a recipe, or something in my local area either. Actually, I very rarely do this.
> But "search" and "getting an answer to a question" are two different things, aren't they?
Google exists, as both a successful enterprise and as a verb, precisely because to most people they are exactly the same thing.
No, this is wrong. People ask what they want to know. Sometimes the best answer is a link. Sometimes it's just an answer. The ability to intuit which is best is what makes products in this space worth making.
Like you, I thought typing questions into Google was wrong for a long time. The times have changed; this is how most people interact with Google, and it really does convey intent to the system better now that we have sufficiently powerful NLP.
That’s okay if your goal is to get an answer to a straightforward question. If, however, your goal is to research a topic, or to find sources for something, or any other scenario where your aim is to read actual web pages, then you want web search, not AI answers. These are two different use cases.
I absolutely agree that it handles natural language questions much better now than when I started using search engines in the late 1990s - in fact it's optimized for this task now, meeting demand where it's at - but a direct answer to a question is often not what I want. For example, I often want to find a page that I remember reading in the past, so that I can re-read or cite it. Or I want more reading material to get a deeper, more nuanced understanding of some topic; things that will provide more context around answers or lead me to generating new questions.
>But "search" and "getting an answer to a question" are two different things, aren't they?
First conceptualization of the "search" were web directories then AltaVista and Google drove the complexity down for the users by providing the actual system which crawls, index and ranks web information. Now cycle will repeat again and we will get Answer Machines aka chat bots which drive the UX complexity for users even more down.
Why would I skim search results links and websites if the "AI" can do it for me. The only reason would be if you don't trust the "AI" and you want the actual links of websites so you can look for useful information by yourself but the majority of people want an instant answer/result to their search query hence Google's old school button "I'm feeling lucky".
Getting an answer to a question is a superset - the answer can be a page.
Sometimes the answer we want is a specific page containing some term, but for most people, most of the time, I'd argue that getting a narrower piece of information is more likely to be valuable and helpful.
May I ask how old you are? I'm 38 and I've been trying hard to break my 10 year-old of the habit of just typing questions into search engines (or telling me to "Ask Google" whenever she asks me a question and I say, "Oh, I don't know").